A self-driving product can prompt itself.
PostHog · 05 May 2026
Ephemeris.
GPT-5.5 quietly takes the steering wheel.
OpenAI's new Instant model is not a flagship. It is the boring middle of the funnel — the one most users will hit on every keystroke — now markedly sharper, with fewer hallucinations and persona controls baked in. The interesting bet is that the median request, not the marquee one, is where the next year of moats are won.
The pixel arrives in ChatGPT.
Buried under the GPT-5.5 launch noise: OpenAI shipped an actual ads SDK. A JavaScript pixel, a server-side conversions API, a programmatic advertiser API. The same primitives that turned Facebook into an ad engine in 2014, now wired to a chat box that 700 million people open every day. Whether you sell on it or not, the attention economy just acquired a new freeway.
Drop-in script for tracking ChatGPT-driven web conversions.
Backend event stream for CAPI-style attribution.
Programmatic campaign creation, bidding, reporting.
A product that prompts itself.
PostHog Code, in beta, is a desktop agent that watches your product data and writes the next ticket. Routine bugs and small UX tweaks get fixed inside the loop; complex work is surfaced as a prioritised backlog with the data already attached. The pitch: stop asking engineers to babysit dashboards — let the dashboard hire the engineer.
[ session-replay ] ─→ [ signal-store ] ─→ [ planner ]
└ "users can't find export → patch UX"
[ planner ] ─→ [ codex.tool ] ─→ [ pr.draft ]
[ pr.draft ] ─→ [ eval.guardrails ] ─→ [ ship | escalate ]
Claude reaches into the art studio.
Anthropic shipped connectors into nine creative tools — Adobe Creative Cloud, Blender, Autodesk Fusion, music DAWs — so Claude can move layers, render comps, and prototype objects without leaving the canvas. The framing is careful: "It can't replace taste or imagination, but it can open up new ways of working — faster and more ambitious ideation."
- Adobe Creative Cloud
- Blender
- Autodesk Fusion
- Music DAWs
- Figma
- Procreate
- After Effects
- Cinema 4D
- Ableton
APPROVAL by phone call.
Zvi reads the tea leaves on a quiet but seismic shift: the White House is moving toward an informal "prior restraint" regime where frontier model releases require an unwritten nod before they ship. He concedes some oversight is reasonable; what he fears is the procedural vacuum — favouritism dressed as safety.
"A good implementation of a prior restraint regime for true frontier model releases, isolated to the biggest models of the leading labs and with formalized procedures that are difficult to abuse, is a good and eventually XXXXXXXX even a necessary thing. I fear that is not what we are going to get."
DRAFT · NO DATEA daily scoreboard for AI on Hacker News.
A small pipeline by Sergei Boldyrev pulls the top 200 HN posts each morning, filters them with an LLM for AI relevance, and asks Gemini to score sentiment per model. The output: a public dashboard where Claude reads as polarised, GPT as quietly tolerated, and Gemini barely registers. Useful as a low-cost early-warning system for a release going sideways.
| # | model | mentions / day | tone | signal |
|---|---|---|---|---|
| 1 | Claude | 142 | + / − | |
| 2 | GPT | 96 | + | |
| 3 | Gemini | 38 | · | |
| 4 | Llama | 21 | + | |
| 5 | Mistral | 9 | · |
That's today.
Six picks from a single news-heavy Tuesday: a default model upgrade, an ads platform that quietly shipped under it, a coding agent that watches its own product, connectors into the creative stack, a policy memo that is not a memo, and a small dashboard scoring the conversation.